Hybrid Approach to Parallel Stochastic Gradient Descent

[Submitted on 27 Jun 2024]

View a PDF of the paper titled Hybrid Approach to Parallel Stochastic Gradient Descent, by Aakash Sudhirbhai Vora and Dhrumil Chetankumar Joshi and Aksh Kantibhai Patel

View PDF
HTML (experimental)

Abstract:Stochastic Gradient Descent is used for large datasets to train models to reduce the training time. On top of that data parallelism is widely used as a method to efficiently train neural networks using multiple worker nodes in parallel. Synchronous and asynchronous approach to data parallelism is used by most systems to train the model in parallel. However, both of them have their drawbacks. We propose a third approach to data parallelism which is a hybrid between synchronous and asynchronous approaches, using both approaches to train the neural network. When the threshold function is selected appropriately to gradually shift all parameter aggregation from asynchronous to synchronous, we show that in a given time period our hybrid approach outperforms both asynchronous and synchronous approaches.